A Fast and Fully-Automated Deep-Learning Approach
Total Page:16
File Type:pdf, Size:1020Kb
www.nature.com/scientificreports OPEN A fast and fully‑automated deep‑learning approach for accurate hemorrhage segmentation and volume quantifcation in non‑contrast whole‑head CT Ali Arab1,10, Betty Chinda2,3,10, George Medvedev4, William Siu5, Hui Guo3,6, Tao Gu3,7, Sylvain Moreno8,9, Ghassan Hamarneh1, Martin Ester1* & Xiaowei Song2,3* This project aimed to develop and evaluate a fast and fully‑automated deep‑learning method applying convolutional neural networks with deep supervision (CNN‑DS) for accurate hematoma segmentation and volume quantifcation in computed tomography (CT) scans. Non‑contrast whole‑head CT scans of 55 patients with hemorrhagic stroke were used. Individual scans were standardized to 64 axial slices of 128 × 128 voxels. Each voxel was annotated independently by experienced raters, generating a binary label of hematoma versus normal brain tissue based on majority voting. The dataset was split randomly into training (n = 45) and testing (n = 10) subsets. A CNN‑DS model was built applying the training data and examined using the testing data. Performance of the CNN‑DS solution was compared with three previously established methods. The CNN‑DS achieved a Dice coefcient score of 0.84 ± 0.06 and recall of 0.83 ± 0.07, higher than patch‑wise U‑Net (< 0.76). CNN‑DS average running time of 0.74 ± 0.07 s was faster than PItcHPERFeCT (> 1412 s) and slice‑based U‑Net (> 12 s). Comparable interrater agreement rates were observed between “method‑human” vs. “human– human” (Cohen’s kappa coefcients > 0.82). The fully automated CNN‑DS approach demonstrated expert‑level accuracy in fast segmentation and quantifcation of hematoma, substantially improving over previous methods. Further research is warranted to test the CNN‑DS solution as a software tool in clinical settings for efective stroke management. Hemorrhagic stroke refers to the loss of brain function due to the accumulation of blood inside the brain arising from compromised cerebral vasculature1,2. It is associated with high death rate and low recovery probability3,4. Over 60 million DALY’s (Disability Adjusted Life Years—a measure of lost time and economic resources) are lost annually to hemorrhagic stroke5, with stroke care costing 34 billion dollars in the United States alone 6. Based on the stroke best practice recommendations, the presence of acute hemorrhagic stroke is confrmed clinically using non-contrast computed tomography (CT) imaging, which visibly distinguishes pathologic blood from normal brain tissue 7. Precision of clinical decision making for hemorrhagic stroke is greatly enhanced by the timely and accurate retrieval and application of information embedded within these scans. Hematoma detection 1School of Computing Science, Simon Fraser University, Burnaby, BC, Canada. 2Department of Biomedical Physiology and Kinesiology, Simon Fraser University, Burnaby, BC, Canada. 3Health Sciences and Innovation, Surrey Memorial Hospital, Fraser Health Authority, Surrey, BC, Canada. 4Division of Neurology, Royal Columbian Hospital, New Westminster, BC, Canada. 5Division of Radiology, Royal Columbian Hospital, New Westminster, BC, Canada. 6Department of Radiology, Tianjin Medical University General Hospital, Tianjin, China. 7Department of Radiology, Beijing Hospital, Beijing, China. 8School of Interactive Arts and Technology, Simon Fraser University, Surrey, BC, Canada. 9Digital Health Circle, Simon Fraser University, Surrey, BC, Canada. 10These authors contributed equally: Ali Arab and Betty Chinda. *email: [email protected]; [email protected] Scientifc Reports | (2020) 10:19389 | https://doi.org/10.1038/s41598-020-76459-7 1 Vol.:(0123456789) www.nature.com/scientificreports/ and assessment of the volume, location, and spread of the bleed are contingent for correct prognostication and treatment outcomes, constituting major determinants of stroke mortality8–11. However, the paucity of fast and accurate hematoma volumetric analysis tools have limited the efective util- ity of CT-based information in guidance of care plans. Current clinical practice for hematoma quantifcation employs primarily the simple ABC (length, width, and height)/2 method, which assumes that the hemorrhage is elliptical and accordingly estimates hemorrhage volume by calculating the volume of the ellipse 12. Research even till recent has consistently shown that manual calculations involved with this clinical standard can be inher- ently time-consuming, expertise dependent, error-prone, and challenging to deal with several complex, large, and irregular hematoma cases11. Given the clinical signifcance, numerous attempts at developing computer-assisted automation tools for hematoma volume have been reported13–16. However, many of the techniques even till recent are either manual, semi-automatic, slow, or sufer from poor accuracy, especially for irregular hematoma shapes and sizes 13–19. Great eforts have been geared towards building automated segmentation tools with use of artifcial intelligence tech- nologies, in particular the promising deep-learning algorithms20,21. Even so, established methods thus far have been largely restricted to handling intracranial hemorrhage and have not been evaluated in clinical settings 20,21. Developing fully-automated deep-learning methods with both high-level of accuracy and efciency for improved clinical management of multiple types of hemorrhagic stroke is holding tremendous research attention 22,23. Previously proposed approaches have attempted to address the problem through optimizing a hematoma mask using level-set techniques for computations of the hematoma surfaces and shapes24–27. Despite the ease to follow complex shapes without the prerequisite of complicated parameterization, the level-set based methods have generally sufered from oversensitivity to the contour initialization, susceptibility to local optima and requiring numerous convergence iterations 24–27. Alternatively, recent research has applied supervised learning methods, where the algorithm can learn the hematoma mapping function under supervision from the ground truth, i.e., previously defned labels indicating the hematoma regions. For example, the well-established PItcHPERFeCT method computes hand-engineered features (i.e., extract useful information from the images using the domain knowledge) for each voxel based on the surrounding patch, and a random forest model is trained to classify each voxel 28. However, despite the high accuracy, PItcHPERFeCT is slow as its feature extraction is computationally very expensive28. As an important aspect of artifcial intelligence, the deep learning approaches have drawn an increasingly great attention in medical imaging analyses, due to its abilities to learn non-linear relationships from the data and to perform automatic feature extraction without using any domain knowledge. Several deep learning applications have tackled intracerebral hemorrhage (ICH) segmentation 17,29,30. Many of the methods are based on the U-Net model, a deep neural network based approach designed for training end-to-end for dense segmentation, i.e., producing voxel-wise labelling 7,29. Although the currently available deep learning methods have shown efective- ness and efciency in segmenting ICH, there are unsolved questions calling for further research. First, most of the previous fndings were based on CT images from a single institute and thus requires generalization testing. Te model’s performance might be afected by slice spacing that can vary between CT scans in multiple center data. Also, the input data to most of the methods were only patches or a small number of consecutive slices of the CT scan, which limited the model’s feld of view while requiring additional inferences to reconstruct the whole-brain images, increasing the time for the error-prone prediction. Here, our team aimed to develop a fast and fully-automated deep-learning oriented approach for accurate hemorrhage segmentation and volume quantifcation on non-contrast whole-head CT31. To achieve this goal, we applied a convolutional neural network with deep supervision (CNN-DS) on a novel dataset collected from three hospitals. Tis modifed U-Net with deep supervision method combined the strengths of CNN in efectively learning visual imagery data with deep supervision to speed up the convergence and reduce over-ftting32, and thereby increasing generalizability of the model. Tis ability is crucial in studying medical images where train- ing sample size is ofen limited by the availability of clinical data and human-expert labeling 32. Tis study marks the frst efort to our knowledge that extends the usability of U-Net with deep supervision for the segmentation and quantifcation of hematoma volume in CT scans. We evaluated the performance of the CNN-DS method in terms of accuracy and efciency, and compared these with those of three recently established well-adopted machine-learning algorithms. We hypothesized that the CNN-DS approach can yield an expert-level of accuracy in segmenting and evaluating hemorrhage volume on head CT images, while being highly time-efcient. Results Training and test cohort. Table 1 provides the characteristics of the patients randomly selected for the training and testing data subsets. Patients in the two subsets showed similar demographics, even though patients in the testing set appeared to be slightly older (t = − 1.6, p = 0.12 two-tailed). Segmentation accuracy, efciency, and reliability. Te CNN-DS method for hematoma quantifca- tion in the training set had a Dice coefcient